Multi-modal large model hallucination uncertainty quantification method and system based on problem condition atomic statement decomposition and evidence calibration

Abstract

The application discloses a multi-modal large model illusion uncertainty quantification method and system based on problem condition atomic declaration decomposition and evidence calibration, constructs declaration level features around image evidence, OCR recognized text evidence, digital evidence, direct verifier support / refutation evidence, multi-view consistency evidence and multi-sampling consistency evidence, and calibrates the model output to have probability meaning of illusion uncertainty through regularization evidence weighting. The application can locate the risk source of specific objects, numbers, OCR text, attributes or spatial relationships by decomposing the visual question answering into problem condition atomic declaration through QACD; can distinguish illusion reasons such as unstable answer, image not supporting, wrong OCR text and digital mismatch; can output threshold rejection, manual review and risk-coverage analysis; and can improve the illusion risk ranking ability and calibration error.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence security, trustworthy multimodal large models, visual question answering (VQA) result evaluation, model output post-processing and risk control, and specifically relates to a method and system for quantifying the illusion uncertainty of multimodal large models based on problem condition atomic declaration decomposition and evidence calibration. Background Technology

[0002] The big model illusion refers to the phenomenon where a large language model fabricates information that it believes to be real or even reasonable or credible, but the content generated by the model is inconsistent with real-world facts or user input.

[0003] Currently, the main solutions to large model illusions fall into the following categories: 1. Posterior confidence calibration techniques: Methods such as temperature scaling, Platt scaling, isometric regression, and Beta calibration map model confidence or single risk scores to probabilities, but they usually rely on response level scores and lack a breakdown of the specific factual statements within the response.

[0004] 2. Sample consistency and self-checking techniques: By estimating uncertainty through multiple sampling responses, clustering, entropy, pairwise differences, and the proportion of unique answers, it is possible to detect unstable responses. However, it usually lacks a stable image evidence verification mechanism, making it difficult to determine whether each visual fact, OCR text, or digital claim is supported by image evidence.

[0005] 3. Text-based Large Language Model Illusion Detection Technology: For example, schemes like "Large Language Model Illusion Detection Based on Multi-Perspective Questioning and Graph Reasoning" typically involve multiple rounds of questioning of the text responses, structuring factual statements into subject-relationship-object-constraint fact atoms, constructing a consistency graph, and calculating the illusion risk score using contradictory edges, drift edges, evidence missing degree, and graph reasoning. The core of this type of scheme is the consistency graph reasoning of textual fact atoms and proactive follow-up questioning.

[0006] 4. Multimodal factual / illusion assessment techniques: Some methods target whole sentences or paragraphs of responses and use validators or external baselines to judge signals such as support, contradiction, and missing evidence, but usually do not form a declaration-level probability calibration process with few sample annotations for the target domain.

[0007] 5. Cross-domain risk calibration techniques: such as calibration, distribution transfer or shrinkage fusion methods based on the distribution of the source domain and a small number of annotations in the target domain, can alleviate domain migration, but their inputs are mostly single response-level risks, making it difficult to utilize structured features such as claim type, evidence alignment, OCR / digit matching.

[0008] In addition, the existing technology also has the following problems: 1. Coarse granularity: Response-level risk can only determine whether the entire answer is suspicious, but cannot pinpoint which specific atomic fact, number, OCR text, entity, or spatial relationship is causing the illusion.

[0009] 2. Insufficient use of evidence: The sample consistency method mainly measures whether the responses are consistent, which can easily confuse model uncertainty with real illusions, and lacks direct alignment of visual, textual and digital evidence.

[0010] 3. Lack of multimodal evidence: Text-based graph reasoning methods mainly rely on consistency, contradictions, and drift relationships between multiple rounds of text responses, making it difficult to directly handle multimodal evidence such as image support, visible text, counting results, target attributes, and spatial relationships in visual question answering.

[0011] 4. Risk scores lack probabilistic meaning: Existing detection methods mostly output weighted risk scores or high, medium and low risk levels. The scores do not necessarily equal the true probability of hallucination, making it difficult to support threshold-based rejection, risk coverage curves and calibration error assessment.

[0012] 5. Unstable target domain migration: When migrating from source domains such as OKVQA to target domains such as TextVQA, single temperature scaling, isometric regression, or Beta calibration may result in risky calibration drift, and overfitting is likely to occur with a small number of target domain annotations.

[0013] 6. Insufficient interpretability: Existing methods are unable to simultaneously output auditable reasons such as "low support, contradiction, missing evidence, inconsistent OCR, inconsistent numbers, and excessive diversity of answers", which is not conducive to threshold setting and manual review after going live.

[0014] 7. Lack of a risk modeling mechanism that integrates ranking discrimination and probability calibration: Existing methods often either only learn the illusory risk ranking score or only perform posterior probability mapping on the existing score. They lack a technical process that learns risk discrimination ability and probability scale simultaneously based on declarative multimodal evidence, making it difficult to directly use the output in risk control scenarios such as threshold rejection and manual review.

[0015] Definitions: QACD (Question-Conditioned Atomic Claim Decomposition): Utilizes the question context to transform the answer to be evaluated into at least one short and faithful atomic claim, and outputs the claim text, source fragment, claim type, verification hint, atomicity score, faithfulness score, and visual verifiability category. The current experimental implementation outputs one claim by default for the main prompt word, and outputs multiple claims when multiple claim decomposition or rule fallback is enabled.

[0016] BIC-lite (Belief Interaction Calibrator-lite): Performs logistic regression calibration with L2 regularization on oriented and normalized claim-level evidence features, learns evidence weights and probabilistic scales, and outputs the probability that the claim is an illusion.

[0017] Unique Answer Ratio: The multi-perspective consistency module groups the responses of each validation perspective for the same declaration by unit_id. After removing leading and trailing whitespace, converting to lowercase, removing English articles, replacing punctuation with spaces, and merging consecutive whitespace in the responses of the group, the number of different values ​​of the non-empty normalized response string is divided by the total number of validation perspective response records in the group. This statistic is used to measure the string diversity of multi-perspective responses and does not use correct answer labels.

[0018] unit_id (Claim Unit Identifier): Used to uniquely identify an atomic claim unit obtained from QACD decomposition; multiple records generated by the same claim under different verification perspectives share the same unit_id, which is used for grouping, summarizing, and subsequent feature table association.

[0019] The Sigmoid function is an S-shaped function that maps any real number to the interval (0,1). Its expression is σ(z)=1 / (1+exp(-z)).

[0020] OCR (Optical Character Recognition): Identifies visible text from an image and outputs OCR-recognized text, which is used to align evidence with visible text in a statement, signage, branding, menus, or screen text.

[0021] EVIDENCE (Evidence Text Field): The textual description of support, rebuttal, or insufficient evidence returned by the multimodal validator; the original string is stored in the validation response record and used for evidence presence marking, text alignment, and original evidence retrieval.

[0022] NUMBER (Number Evidence Field): Numerical content such as quantity, price, and number related to the claim, returned or parsed by the multimodal validator, is used to verify the numerical consistency with the set of numbers in the claim.

[0023] feature_names (Feature Name List): Represents the fixed names and order of each feature column in the BIC-lite input matrix; the k-th name uniquely corresponds to the k-th column and the k-th learned weight in the matrix, which is used to ensure that features, weights and interpretation results are aligned according to the same index. Summary of the Invention

[0024] To address the aforementioned issues, this invention discloses a method and system for quantifying the uncertainty of multimodal large model illusions based on problem-conditional atomic declaration decomposition and evidence calibration.

[0025] To achieve the above objectives, the technical solution of the present invention is as follows: A method for quantifying the illusion uncertainty of a multimodal large model based on problem-conditional atomic declaration decomposition and evidence calibration includes the following steps: S1. Input and Preprocessing: Input image I and question Q to the multimodal large model to be evaluated. The multimodal large model to be evaluated outputs the answer to be evaluated, A. Then, the multimodal large model to be evaluated generates a set G={A(1),A(2),…,A(M)} of M sampled answers for the input image I and question Q using different random sampling seeds. I, Q, A and G constitute the data to be processed. The data to be processed is normalized and preprocessed to obtain a unified input representation. A(M) represents the answer obtained by the multimodal large model to be evaluated through the Mth random sampling inference of image I and question Q. S2. Problem Condition Atomic Statement Decomposition: Using QACD, the answer to be evaluated, A, is transformed into a set of at least one short and faithful atomic statement. ; where the i-th declaration It includes claim text, source fragment, claim type, verification prompt, atomicity score, fidelity score, and visual verifiability category, forming a multimodal claim unit that can be directly verified by images, OCR-recognized text, numbers, and verifiers; N is the number of atomic claims; S3. Direct Evidence Verification: Combining Image I, Question Q, and Statement The declaration text and validation prompts are input into the multimodal validator to obtain the structured validator output; S4. Multimodal Evidence Alignment: Align the feature categories in the structured validator output to obtain the declaration-level evidence vector. ; S5. Regularized Evidence Weighting Calibration: Adjusting the declaration-level evidence vector... Input a pre-trained BIC-lite calibrator to learn the mapping from evidence to hallucination probabilities, and obtain the declaration-level hallucination probability. ; S6, Risk Output: Output the illusion probability for each claim. Answer-level aggregation risk And explanatory evidence.

[0026] Further improvements, the specific steps of step S1 are as follows: The evaluation response A is generated by the multimodal large model to be evaluated in response to the input image I and question Q. Then, a set G = {A(1),A(2),…,A(M)} is formed by M sampled responses generated by the multimodal large model to be evaluated under the same image I and question Q using different random sampling seeds. The image I, question Q, evaluation response A and set G form the data to be processed. The steps for normalizing the data to be processed are as follows: Clean up redundant spaces, punctuation marks, and prompt prefixes in the image path or image I identifier, the text of question Q, the text of the answer to be evaluated A, and the text of set G; unify the field format; retain the sample number, dataset identifier, and evaluation label interface; and establish a unified input representation consisting of image I, question Q, answer to be evaluated A, and multiple sampled answer set G.

[0027] In a further improvement, in step S3, the multimodal verifier is a visual language model with the ability to reason with both text and images, or a reasoning instance that has the same parameter base as the model to be evaluated but runs with independent verification prompts. The multimodal validator executes the direct support perspective, visual evidence perspective, OCR reading perspective, and digital verification perspective respectively, to obtain a structured validator output, which includes the average support probability of each verification perspective. The average probability of rebuttal from each verification perspective Uncertain probability Low support probability Matching probability Low matching probability The visual language model includes LLaVA-1.5-13B, which includes confidence level, evidence presence marker, evidence text field EVIDENCE, OCR recognition text field and number field NUMBER.

[0028] In a further improvement, in step S4, each feature category is expanded into the actual scalar input features of BIC-lite according to the pre-determined column names and column order in feature_names. The declaration-level evidence vector includes, but is not limited to, the following: : in, Decompose low-quality risk feature groups for QACD. ,in, Because of its low atomicity, For low loyalty, Low decomposition confidence; For the declaration Type one-hot encoding or type embedding; This represents the average of the uncertainty probabilities from each verification perspective; This represents the average probability of support for each verification perspective; Low support probability =1- ; Low matching probability =1- ; Alignment score for text evidence, i.e. The larger of the sequence similarity and keyword coverage between the evidence text and the statement text; The OCR evidence alignment score, i.e. The larger of the sequence similarity and keyword coverage between the OCR-recognized text and the declaration text; Alignment score for digital evidence: 1 for a declaration containing numbers and matching the set of numbers, 0 for a declaration containing numbers but not matching, and 0.5 for a declaration containing no numbers. It is a multi-perspective statistical feature vector, including support missing rate, refutation rate, perspective disagreement degree, and unique answer ratio; To address the inconsistency feature from multiple sampling: Cluster the answers in set G according to the rules that, after normalization preprocessing, the answers are completely identical, mutually inclusive, or have a Jaccard similarity of not less than 0.72. Let the maximum number of samples per cluster be... ,but Take 1 if there is no sampled response; To support the composition of a multi-perspective statistical feature vector based on missing rate, refutation rate, perspective disagreement, and unique response ratio in a fixed order: The method for obtaining the results is as follows: Calculate the validation perspective response for each of the V validation perspectives: Support Missing Rate Refutation rate and the degree of difference in perspective The proportion of unique answers ;in, This indicates that the i-th statement is in the... v Support score from a verification perspective This represents the maximum score supported by the same statement across V validation perspectives. The minimum support score for the same statement across V validation perspectives is represented by the difference between the two values, which is the range of the support scores. The range is 0 to 1. A difference of 0 indicates that the support scores are consistent across all validation perspectives, while a larger difference indicates stronger divergence among the perspectives. To record the number of values ​​for the non-empty, normalized ANSWER string, To verify the total number of viewpoint response records.

[0029] Further improvements, the specific steps of step S5 are as follows: Declarative-level evidence vectors Input the trained BIC-lite calibrator to learn the mapping from evidence to hallucination probability: First, calculate the log odds of the declared risk using a regularized evidence-weighted summation algorithm. : in, Let be the log-odds of the risk of the i-th statement. For bias terms, The feature dimension of the declaration-level evidence vector. For the learnable weights of the k-th type of evidence, For the k-th standardized evidentiary feature of the i-th statement, Let T be a column vector consisting of all learnable weights, and let T denote the matrix transpose. Then, the declaration-level hallucination probability is obtained through a logistic sigmoid function mapping. : in, Indicates the first The statement is an illusion. For binary labels, 1 indicates hallucination and 0 indicates non-hallucination; () represents a logical S-shaped function. () represents an exponential function with the natural constant e as its base.

[0030] Further improvements involve training the BIC-lite calibrator to obtain a trained BIC-lite calibrator, with the loss function used during training... as follows: in, The target domain has a labeled set of calibration samples. The L2 regularization coefficient is greater than 0; training to When the minimum value is reached, a trained BIC-lite calibrator is obtained.

[0031] A further improvement is made to the BIC-lite calibrator, which employs declaratively conditional gated feature fusion: in, The log odds of declared risk obtained after inputting gated interaction features into the BIC-lite calibrator; For bias terms; K This refers to the dimension of the input features after expansion according to the fixed column names and column order of feature_names; For the first k Learnable weights of standardized features; For the i-th statement targeting the i-th k The gating condition scalar of the evidence class is obtained by multiplying one or more of the following: the statement type indicator, the OCR or digital verification intent, the validator confidence or its complement, and the evidence availability. For the k-th gated interaction feature of the i-th declaration, The basic risk characteristics are pre-bound to the k-th gating condition.

[0032] A further improvement is made in step S6, which outputs the illusion probability for each claim. Risk ranking, answer-level aggregated risk And explanatory evidence; On the labeled calibration samples in the target domain, the illusion probability of each calibration statement is used as a risk score. Empirical quantiles are calculated and deduplicated at quantile levels 0, 1 / 200, 2 / 200, ..., 1 to form a candidate threshold set. For each candidate threshold τ, calibration statements with a risk score no higher than τ are considered accepted samples. Coverage and the empirical illusion rate of accepted samples are calculated. The maximum threshold τ, where the empirical illusion rate is no higher than a preset risk upper limit α and the coverage is no lower than a preset minimum coverage, is selected. If no candidate threshold meets the conditions, the minimum risk score is used as the threshold. Statements with a probability exceeding τ are marked as objects for manual review, and risks are aggregated at the response level. When the response risk threshold is exceeded by a separate determination based on the same quantile and risk constraints on the response-level labeled calibration sample, a rejection recommendation is given. Answer-level aggregation risk Calculated as the arithmetic mean of the N claimed hallucination probabilities within the same sample: Or answer the question about aggregation risk Take the maximum declared risk Or answer the question about aggregation risk Weighted average based on the number of normalized characters in the declaration: Among them, the importance weight is declared. ; The number of characters after whitespace and punctuation normalization of the i-th declaration. This represents the number of characters after performing the same normalization on the j-th declaration; The interpretable evidence includes the structured validation fields corresponding to the statement, the evidence alignment score, and the weighted contribution of standardized evidence features; Structured validation fields include SUPPORT, CONTRADICT, MATCH, EVIDENCE, OCR, and NUMBER fields; evidence alignment scores include text, OCR, and number alignment scores. The single-weighted contribution of the standardized evidentiary features according to Calculate; where, For the k-th standardized evidentiary feature of the i-th statement, These are the corresponding global learning weights; ranked by influence strength. Sort the evidence features from largest to smallest, and select the top L standardized evidence features, their corresponding global learning weights, and individual weighted contributions as interpretable evidence. The structured verification fields are obtained by returning and parsing the verification perspectives in step S3 according to a fixed response format. The structured verification fields include support probability, rebuttal probability, uncertainty probability, matching probability, confidence level, evidence text field EVIDENCE, OCR recognized text field, number field NUMBER, and evidence presence marker. The evidence alignment score includes the text evidence alignment score. OCR evidence alignment score Alignment score with digital evidence .

[0033] A multimodal large model illusion uncertainty quantification system based on problem conditional atomic declaration decomposition and evidence calibration is provided to perform the aforementioned multimodal large model illusion uncertainty quantification method based on problem conditional atomic declaration decomposition and evidence calibration.

[0034] Further improvements include an input and preprocessing module, a problem condition atomic declaration decomposition module, a direct evidence verification module, a multimodal evidence alignment module, a regularized evidence weighting calibration module, and a risk output module; The input and preprocessing module is used to receive image I, question Q, answer A to be evaluated, and multiple sampled answer set G and obtain a unified input representation; the question condition atomic declaration decomposition module is used to decompose the answer A to be evaluated in the unified input representation into question condition atomic declarations to obtain multimodal declaration units including declaration text, source fragment, declaration type, verification prompt and decomposition quality score; The direct evidence verification module is used to generate multi-perspective verification input based on the verification prompts in the multimodal declaration unit, and call the multimodal verifier to obtain a structured verifier output consisting of support, rebuttal, uncertainty, matching, confidence, evidence presence marker, evidence text, OCR-recognized text and numeric fields; The multimodal evidence alignment module is used to align the structured validator output with the declaration text using text, OCR, and numbers, construct multi-view statistical features and multiple sampling inconsistency features, and obtain the declaration-level evidence vector. ; The regularized evidence weighting calibration module is used to adjust the declaration-level evidence vector. Input the BIC-lite calibrator to learn the mapping from evidence to hallucination probability, and obtain the hallucination probability for each claim. ; The risk output module is used to output the illusion probability of each claim. Answer-level aggregation risk Explainable evidence, including the structured verification field corresponding to the statement, the evidence alignment score, and the single-weighted contribution of standardized evidence features; risk ranking is performed according to the statement-level illusion probability, and a candidate threshold set is formed based on 201 equidistant empirical quantiles of the target domain labeled calibration samples; statement-level risk thresholds and response-level risk thresholds that meet the upper risk limit and minimum coverage constraints are determined respectively, and suggestions for manual review or rejection are given.

[0035] The advantages of this invention are as follows: 1. Declaration-level positioning effect: By breaking down visual question answering into question condition atomic declarations through QACD, the risk source can be located for specific objects, numbers, OCR text, attributes or spatial relationships, rather than just giving the entire answer an unexplainable score.

[0036] 2. Evidence alignment effect: By combining direct image evidence, OCR text evidence, digit consistency evidence, validator support / rebuttal probability, and multiple sampling consistency features, it can distinguish different illusionary causes such as "unstable response", "image not supported", "text misreading" and "digit mismatch".

[0037] 3. Probabilistic calibration effect: BIC-lite learns evidence weights and probabilistic scales on labeled calibration samples in the target domain, so that the output is not only a hallucination ranking score, but also has a probabilistic meaning that can be used for threshold rejection, manual review and risk-coverage analysis.

[0038] 4. Difference from text graph reasoning methods: This invention does not rely on the subject-relationship-object-constraint quadruple and consistency graph contradictory path search, but instead completes the quantification of illusion uncertainty by focusing on multimodal visual evidence alignment and declaration-level probability calibration.

[0039] 5. Experimental verification results: With 300 labeled calibration samples in the TextVQA target domain, the QACD-v3+BIC-lite of the present invention outperforms the comparative methods listed in Table 1 in terms of hallucination risk ranking ability and calibration error. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0041] like Figure 1 As shown, this invention provides a declaration-level illusion uncertainty quantification method for visual question answering generation results, which consists of five core parts: "atom declaration decomposition of question conditions, direct evidence verification, multimodal evidence alignment, regularized evidence weighting calibration, and risk output".

[0042] 1. Input and Preprocessing: Receive image ,question The evaluation response generated by the multimodal large model to be evaluated for image I and question Q. And optionally, a set of multiple sampled responses generated using different random sampling seeds under the same image I and question Q conditions. The system cleans up redundant spaces, punctuation marks, and prompt prefixes in image labels, question text, and answer text, standardizes field formats, and retains sample numbers, dataset identifiers, and evaluation label interfaces.

[0043] This step inputs image I and question Q into the multimodal large model to be evaluated to obtain the answer A to be evaluated. When multiple sampling is enabled, the random sampling seed is changed for the same image I and question Q to obtain set G. Then, a unified input representation is established as the input data for statement decomposition, evidence verification and evidence calibration.

[0044] 2. Problem condition atomic declaration decomposition: Using QACD to decompose the answer Transform into a set consisting of at least one short and faithful atomic statement. Each statement This includes the claim text, source fragment, claim type, verification hint, atomicity score, fidelity score, and visual verifiability category. This atomic claim differs from the quadruple fact atoms in text graph reasoning; its purpose is to form multimodal claim units that can be directly verified by images, OCR-recognized text, numbers, and verifiers.

[0045] This step uses the QACD atomic declaration decomposition algorithm to output a set of declarations that can be verified by images, OCR-recognized text, numbers, and validators, while retaining quality features such as declaration type, atomicity, fidelity, and visual verifiability. The i-th declaration adopts low atomicity, low fidelity, and low decomposition confidence as quality features consistent with the risk direction, and each component is taken as 1 minus the corresponding score.

[0046] 3. Direct evidence verification: for each claim The system reads the verification hints output by QACD and feeds Image I, Question Q, Declaration Text, and verification hints into a multimodal validator capable of joint image-text reasoning. The validator can be LLaVA-1.5-13B, or an inference instance with the same parameter base as the model to be evaluated but running with independent verification hints. It performs direct support, visual evidence, OCR reading, and digital verification perspectives to obtain the support probability. Probability of refutation Uncertain probability Low support probability Matching probability and low matching probability And extract the evidence text field ( ), OCR recognition of text fields ( ) and numeric fields ( The output of structured validators such as )

[0047] The direct evidence verification module takes images, questions, statement text, and verification prompts as verification inputs and performs direct support, visual evidence, OCR reading, and digital verification perspectives. The structured outputs of each perspective are parsed into support, rebuttal, uncertainty, matching, confidence, and evidence availability signals, and participate in weighted calibration in subsequent evidence vectors.

[0048] 4. Multimodal Evidence Alignment: For the i-th claim text, align the evidence text field ( After normalization, the character sequence similarity and non-stop word coverage are calculated, and the larger value is taken as the evidence text alignment score; the OCR-recognized text field ( Match the visible text, signage, branding, menu, or screen text in the declaration in the same way; match numeric fields ( The numerical consistency between the quantity, price, or number stated in the declaration is checked; This represents the support score for the i-th statement under the v-th validation perspective. The support missing rate, rebuttal rate, and perspective divergence are calculated based on the scores from all perspectives. The multi-perspective consistency module groups the validation perspective responses for the same statement by unit_id. Within each group, the response strings are sequentially subjected to whitespace removal (leading and trailing whitespace), conversion to lowercase, removal of English articles, replacement of punctuation with spaces, and merging of consecutive whitespace. The unique response ratio is obtained by dividing the number of distinct values ​​of the non-empty normalized response string by the total number of validation perspective response records in that group. These four items are then combined in a fixed order to form a multi-perspective statistical feature vector. The unique response ratio does not use correct answer labels; it is only used to measure the string diversity of multi-perspective responses.

[0049] This step lowercases, normalizes whitespace and punctuation in the text, calculates sequence similarity, keyword coverage, and digit set matching results, and uses low atomicity, low fidelity, low decomposition confidence, claim type, support / refutation / uncertainty probability, low support probability, low matching probability, evidence text alignment, OCR evidence alignment, digit evidence alignment, multi-view statistical features, and multiple sampling inconsistency as feature categories. These are then expanded into the actual scalar input features of BIC-lite according to the pre-determined column names and column order in feature_names. Continuous features are standardized using the mean and standard deviation of the labeled calibration set before BIC-lite training. in, The low-quality risk feature group for QACD decomposition consists of low atomicity, low fidelity, and low decomposition confidence. Each component is 1 minus the corresponding score, and expanded into three scalar features according to feature_names. The average of the three is taken when constructing quality interaction features. To declare a type for one-hot encoding or type embedding; To support the probability mean; The mean of the rebuttal probability; The mean of the uncertain probability; Low support probability ; Low matching probability; Alignment score for evidence text; The OCR evidence alignment score; Align scores for digital evidence; It is a multi-perspective statistical feature vector composed of support missing rate, refutation rate, perspective divergence degree and unique answer ratio, and expanded in a fixed order. The inconsistency feature of multiple sampling is 1 minus the proportion of the largest cluster in the multiple sampling responses.

[0050] 5. Regularized Evidence Weighting and Calibration: Combining the above features into a declaration-level evidence vector. The continuous features are standardized using the mean and standard deviation of the calibration samples labeled in the target domain, and then the mapping from evidence to hallucination probability is learned through BIC-lite. This step learns evidence weights on the calibration samples labeled in the target domain and suppresses overfitting through an L2 regularization term.

[0051] This step uses a regularized evidence-weighted summation algorithm to first calculate the log odds of the declared risk. : Then, the declaration-level hallucination probability is obtained through logistic sigmoid function mapping: in, It is the first Learnable weights of class evidence For bias terms, Indicates the first The statement is an illusion. This indicates that the i-th declaration is not an illusion.

[0052] To ensure that the weighted summation result has target domain probability calibration significance, this step further uses the target domain of BIC-lite to train the loss function: in, The target domain has a labeled set of calibration samples. The L2 regularization coefficient is greater than 0. This loss function enables the calibrator to learn both evidence weights and probability scales simultaneously.

[0053] In an optional implementation, step 5 may also use claim-condition gating feature fusion to construct a gating condition scalar based on the claim type, OCR or digital verification intent, validator confidence, and evidence availability. This scalar is then multiplied by the corresponding risk features to obtain gating interaction features. Finally, the basic evidence features, gating condition scalar, and gating interaction features are input into the BIC-lite calibrator in the fixed column order of feature_names. in, The gating condition scalar for the i-th statement against the k-th type of evidence is obtained by multiplying one or more of the following: statement type indicator, OCR or digital verification intent, validator confidence or its complement, and evidence availability. For the k-th gated interaction feature of the i-th declaration, The corresponding basic risk features are used; the basic evidence features, gating condition scalars, and gating interaction features are input into the BIC-lite calibrator in the order of fixed columns in feature_names, and BIC-lite learns the global weights of each feature.

[0054] 6. Risk Output: Output the probability of illusion for each claim. Answer-level aggregation risk Explainable evidence; Explainable evidence consists of the weighted contribution of structured verification fields, evidence alignment scores, and standardized evidence features, and is obtained through index mapping between claims and evidence vectors; Risk ranking is formed from high to low according to the claim-level illusion probability; A candidate threshold set is constructed on 201 equidistant empirical quantiles of the target domain labeled calibration samples, and the maximum threshold that meets the preset risk upper limit and minimum coverage is selected. Claims exceeding the threshold are sent for manual review, and a rejection suggestion is given when the aggregated risk at the response level exceeds the response-level risk threshold determined separately on the response-level labeled calibration samples according to the same rules.

[0055] This step uses a declaration-level risk aggregation algorithm to obtain the answer-level illusion uncertainty; the current code implementation calculates the answer-level aggregated risk by the arithmetic mean of the probabilities of each declared illusion within the same sample: In an alternative implementation, the answer-level aggregate risk can be taken as the maximum claim risk or a weighted average based on the number of normalized characters in the claim: Among them, the importance weight is declared. ; The number of characters after whitespace and punctuation normalization of the i-th declaration. This represents the number of characters after performing the same normalization on the j-th declaration.

[0056] The experimental procedure is as follows: 1. Data Construction: Select TextVQA target domain samples, construct a declaration-level illusion calibration set and test set using 300 labeled calibration samples, and summarize the test results for 5 random seeds.

[0057] 2. Declaration decomposition: Run QACD on the answer to be evaluated to generate atomic declarations of question conditions, declaration types, verification hints, and visual verifiability markers.

[0058] 3. Evidence Extraction: Extract image support, rebuttal, OCR alignment, digit alignment, missing evidence, multi-view disagreement, and multiple sampling inconsistency features for each atomic claim. The FaithScore-style and UNIHD-style proxy methods in the table use a local LLaVA model as a multimodal validator or proxy scoring module in their implementation.

[0059] 4. Calibration and Comparison: The method of this invention uses BIC-lite to learn evidence weighted mapping; the comparison methods include sampling consistency + isometric regression, FaithScore surrogate method + isometric regression, SelfCheckGPT surrogate method + isometric regression and UNIHD surrogate method + isometric regression.

[0060] 5. Indicator Calculation: The area under the receiver operating characteristic curve (AUROC) was used to measure the ranking and discrimination ability of hallucination claims, and the risk expectation calibration error (Risk-ECE) was used to measure the consistency between the predicted hallucination probability and the actual hallucination rate.

[0061] As shown in Table 1, the AUROC of QACD-v3+BIC-lite is 0.801700, which is higher than other comparison methods in the table; the Risk-ECE is 0.043811, which is lower than other comparison methods in the table, indicating that it has both strong ability to distinguish hallucination ranking and good probability calibration effect.

[0062] Table 1. Comparative experimental results of TextVQA target domain using 300 labeled calibration samples (mean of 5 random seeds) Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for quantifying the illusion uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration, characterized in that, Includes the following steps: S1. Input and Preprocessing: Input image I and question Q to the multimodal large model to be evaluated. The multimodal large model to be evaluated outputs the answer to be evaluated, A. Then, the multimodal large model to be evaluated generates a set G={A(1),A(2),…,A(M)} of M sampled answers for the input image I and question Q using different random sampling seeds. I, Q, A and G constitute the data to be processed. The data to be processed is normalized and preprocessed to obtain a unified input representation. A(M) represents the answer obtained by the multimodal large model to be evaluated through the Mth random sampling inference of image I and question Q. S2. Problem Condition Atomic Statement Decomposition: Using QACD, the answer to be evaluated, A, is transformed into a set of at least one short and faithful atomic statement. ; where the i-th declaration It includes claim text, source fragment, claim type, verification prompt, atomicity score, fidelity score, and visual verifiability category, forming a multimodal claim unit that can be directly verified by images, OCR-recognized text, numbers, and verifiers; N is the number of atomic claims; S3. Direct Evidence Verification: Combining Image I, Question Q, and Statement The declaration text and validation prompts are input into the multimodal validator to obtain the structured validator output; S4. Multimodal Evidence Alignment: Align the feature categories in the structured validator output to obtain the declaration-level evidence vector. ; S5. Regularized Evidence Weighting Calibration: Adjusting the declaration-level evidence vector... Input a pre-trained BIC-lite calibrator to learn the mapping from evidence to hallucination probabilities, and obtain the declaration-level hallucination probability. ; S6, Risk Output: Output the illusion probability for each claim. Answer-level aggregation risk And explanatory evidence.

2. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 1, characterized in that, The specific steps of step S1 are as follows: The evaluation response A is generated by the multimodal large model to be evaluated in response to the input image I and question Q. Then, a set G = {A(1),A(2),…,A(M)} is formed by M sampled responses generated by the multimodal large model to be evaluated under the same image I and question Q using different random sampling seeds. The image I, question Q, evaluation response A and set G form the data to be processed. The steps for normalizing the data to be processed are as follows: Clean up redundant spaces, punctuation marks, and prompt prefixes in the image path or image I identifier, the text of question Q, the text of the answer to be evaluated A, and the text of set G; unify the field format; retain the sample number, dataset identifier, and evaluation label interface; and establish a unified input representation consisting of image I, question Q, answer to be evaluated A, and multiple sampled answer set G.

3. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 1, characterized in that, In step S3, the multimodal verifier is a visual language model with the ability to perform joint text-image reasoning or a reasoning instance that has the same parameter base as the model to be evaluated but runs with independent verification prompts. The multimodal validator executes the direct support perspective, visual evidence perspective, OCR reading perspective, and digital verification perspective respectively, to obtain a structured validator output, which includes the average support probability of each verification perspective. The average probability of rebuttal from each verification perspective Uncertain probability Low support probability Matching probability Low matching probability The visual language model includes LLaVA-1.5-13B, which includes confidence level, evidence presence marker, evidence text field EVIDENCE, OCR recognition text field and number field NUMBER.

4. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 3, characterized in that, In step S4, each feature category is expanded into the actual scalar input features of BIC-lite according to the pre-determined column names and column order in feature_names. The declaration-level evidence vector includes, but is not limited to, the following: : in, Decompose low-quality risk feature groups for QACD. ,in, Because of its low atomicity, For low loyalty, Low decomposition confidence; For the declaration Type one-hot encoding or type embedding; This represents the average of the uncertainty probabilities from each verification perspective; This represents the average probability of support for each verification perspective; Low support probability =1- ; Low matching probability =1- ; Alignment score for text evidence, i.e. The larger of the sequence similarity and keyword coverage between the evidence text and the statement text; The OCR evidence alignment score, i.e. The larger of the sequence similarity and keyword coverage between the OCR-recognized text and the declaration text; Alignment score for digital evidence: 1 for a declaration containing numbers and matching the set of numbers, 0 for a declaration containing numbers but not matching, and 0.5 for a declaration containing no numbers. It is a multi-perspective statistical feature vector, including support missing rate, refutation rate, perspective disagreement degree and unique answer ratio; To address the inconsistency feature from multiple sampling: Cluster the answers in set G according to the rules that, after normalization preprocessing, the answers are completely identical, mutually inclusive, or have a Jaccard similarity of not less than 0.

72. Let the maximum number of samples per cluster be... ,but If no sampled response is received, take 1. To support the composition of a multi-perspective statistical feature vector based on missing rate, refutation rate, perspective disagreement, and unique response ratio in a fixed order: The method for obtaining the results is as follows: Calculate the validation perspective response for each of the V validation perspectives: Support Missing Rate Refutation rate and the degree of difference in perspective The proportion of unique answers ;in, This indicates that the i-th statement is in the... v Support score from a verification perspective This represents the maximum score supported by the same statement across V validation perspectives. The minimum support score for the same statement across V validation perspectives is represented by the difference between the two values, which is the range of the support scores. The range is 0 to 1. A difference of 0 indicates that the support scores are consistent across all validation perspectives, while a larger difference indicates stronger divergence among the perspectives. To record the number of values ​​for the non-empty, normalized ANSWER string, To verify the total number of viewpoint response records.

5. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 1, characterized in that, The specific steps of step S5 are as follows: Declarative-level evidence vectors Input the trained BIC-lite calibrator to learn the mapping from evidence to hallucination probability: First, calculate the log odds of the declared risk using a regularized evidence-weighted summation algorithm. : in, Let be the log-odds of the risk of the i-th statement. For bias terms, The feature dimension of the declaration-level evidence vector. For the learnable weights of the k-th type of evidence, For the k-th standardized evidentiary feature of the i-th statement, Let T be a column vector consisting of all learnable weights, and let T denote the matrix transpose. Then, the declaration-level hallucination probability is obtained through a logistic sigmoid function mapping. : in, Indicates the first The statement is an illusion. For binary labels, 1 indicates hallucination and 0 indicates non-hallucination; () represents a logical sigmoid function. () represents an exponential function with the natural constant e as its base.

6. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 1, characterized in that, The trained BIC-lite calibrator is obtained by training the BIC-lite calibrator, and the loss function during training is... as follows: in, The target domain has a labeled set of calibration samples. The L2 regularization coefficient is greater than 0; training to When the minimum value is reached, a trained BIC-lite calibrator is obtained.

7. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 1, characterized in that, The BIC-lite calibrator employs declaratively gated feature fusion: in, The log odds of declared risk obtained after inputting gated interaction features into the BIC-lite calibrator; K represents the bias term; K is the dimension of the input features after expansion according to the fixed column names and column order of feature_names. For the first k Learnable weights of a standardized feature; The gating condition scalar for the i-th statement against the k-th type of evidence is obtained by multiplying one or more of the following: statement type indicator, OCR or digital verification intent, validator confidence or its complement, and evidence availability. For the k-th gated interaction feature of the i-th declaration, The basic risk characteristics are pre-bound to the k-th gating condition.

8. The method for quantifying the uncertainty of multimodal large models based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 4, characterized in that, In step S6, the hallucination probability of each claim is output. Risk ranking, answer-level aggregated risk And explanatory evidence; On the labeled calibration samples in the target domain, the hallucination probability of each calibration statement is used as a risk score. Empirical quantiles are calculated and deduplicated at quantile levels 0, 1 / 200, 2 / 200, ..., 1 to form a candidate threshold set. For each candidate threshold τ, calibration statements with a risk score not higher than τ are considered accepted samples. The coverage and empirical hallucination rate of accepted samples are calculated. The maximum threshold τ with an empirical hallucination rate not higher than a preset risk upper limit α and a coverage not lower than a preset minimum coverage rate is selected. If no candidate threshold meets the conditions, the minimum risk score is taken as the threshold. Declarations with a probability exceeding τ are marked as objects for manual review, and risks are aggregated at the response level. When the response risk threshold is exceeded by a separate determination based on the same quantile and risk constraints on the response-level labeled calibration sample, a rejection recommendation is given. Answer-level aggregation risk Calculated as the arithmetic mean of the N claimed hallucination probabilities within the same sample: Or answer the question about aggregation risk Take the maximum declared risk Or answer the question about aggregation risk Weighted average based on the number of normalized characters in the declaration: Among them, the importance weight is declared. ; The number of characters after whitespace and punctuation normalization of the i-th declaration. This represents the number of characters after performing the same normalization on the j-th declaration; The interpretable evidence includes the structured validation fields corresponding to the statement, the evidence alignment score, and the weighted contribution of standardized evidence features; Structured validation fields include SUPPORT, CONTRADICT, MATCH, EVIDENCE, OCR, and NUMBER fields; evidence alignment scores include text, OCR, and number alignment scores. The single-weighted contribution of the standardized evidentiary features according to Calculate; where, For the k-th standardized evidentiary feature of the i-th statement, These are the corresponding global learning weights; ranked by influence strength. Sort the evidence features from largest to smallest, and select the top L standardized evidence features, their corresponding global learning weights, and individual weighted contributions as interpretable evidence. The structured verification fields are obtained by returning and parsing the verification perspectives in step S3 according to a fixed response format. The structured verification fields include support probability, rebuttal probability, uncertainty probability, matching probability, confidence level, evidence text field EVIDENCE, OCR recognized text field, number field NUMBER, and evidence presence marker. The evidence alignment score includes the text evidence alignment score. OCR evidence alignment score Alignment score with digital evidence .

9. A multimodal large model illusion uncertainty quantification system based on problem-conditional atomic declaration decomposition and evidence calibration, characterized in that, This method is used to perform the multimodal large model illusion uncertainty quantification method based on problem-conditional atomic declaration decomposition and evidence calibration as described in any one of claims 1-8.

10. The multimodal large model illusion uncertainty quantification system based on problem-conditional atomic declaration decomposition and evidence calibration as described in claim 9, characterized in that, It includes an input and preprocessing module, a problem condition atomic declaration decomposition module, a direct evidence verification module, a multimodal evidence alignment module, a regularized evidence weighting calibration module, and a risk output module; The input and preprocessing module is used to receive image I, question Q, answer A to be evaluated, and multiple sampled answer set G and obtain a unified input representation; the question condition atomic declaration decomposition module is used to decompose the answer A to be evaluated in the unified input representation into question condition atomic declarations to obtain multimodal declaration units including declaration text, source fragment, declaration type, verification prompt and decomposition quality score; The direct evidence verification module is used to generate multi-perspective verification input based on the verification prompts in the multimodal declaration unit, and call the multimodal verifier to obtain a structured verifier output consisting of support, rebuttal, uncertainty, matching, confidence, evidence presence marker, evidence text, OCR-recognized text and numeric fields; The multimodal evidence alignment module is used to align the structured validator output with the declaration text using text, OCR, and numbers, construct multi-view statistical features and multiple sampling inconsistency features, and obtain the declaration-level evidence vector. ; The regularized evidence weighting calibration module is used to adjust the declaration-level evidence vector. Input the BIC-lite calibrator to learn the mapping from evidence to hallucination probability, and obtain the hallucination probability for each claim. ; The risk output module is used to output the illusion probability of each claim. Answer-level aggregation risk Explainable evidence, including the structured verification field corresponding to the statement, the evidence alignment score, and the single-item weighted contribution of standardized evidence features; risk ranking is performed according to the statement-level illusion probability, and a candidate threshold set is formed based on 201 equidistant empirical quantiles of the target domain labeled calibration samples; statement-level risk thresholds and response-level risk thresholds that meet the upper risk limit and minimum coverage constraints are determined respectively, and suggestions for manual review or rejection are given.

Images